Hierarchical product quantization for effective feature indexing

Feature indexing is a critical technique for addressing real-time image matching and retrieval. In this work, we propose a novel quantization method that is capable of creating highly accurate quantized codes for a given feature database. Differing from many quantization techniques in the literature (typically, product quantization based methods), the proposed method is designed to reshape the feature vectors so that close points are placed into a small sub-space. To this aim, a hierarchical product quantization method is presented. In its essence, the first level of quantization aims at reordering the dimensions so as to exploit better the correlation among subspaces. The second level of quantization is then invoked to create a sub-quantizer for the points contained in each sub-space. To validate the proposed method, various experiments have been conducted, demonstrating quite impressive performance when compared with other state-of-the-art methods.